Llama.cpp Mac Beats CPU, Hybrid GPU

The takeaway is less “Mac always wins” than “memory bandwidth and prompt shape dominate local inference economics.” For short, chatty workloads, Apple silicon looks unusually strong; for long-context batchy jobs, offload-heavy GPU rigs still have a lane.

• –Test setup used `-c 260000`, `--jinja`, and `--no-mmap`, so this is a high-context local-inference benchmark, not a toy run
• –The result favors Mac on smaller prompts, which is exactly where unified memory can outperform awkward CPU/GPU shuffling
• –GPU+CPU offload only wins when the prompt is several thousand tokens and the completion is comparatively short
• –MX quants were excluded, so the comparison stays apples-to-apples on accuracy rather than chasing the fastest possible speed
• –Treat this as a configuration note, not a universal verdict; quant type, context length, and backend kernels can easily reshuffle the rankings